A central component of every decision support system is a performance prediction model. Often, empirical data is used to describe a performance measure as a function of factors affecting performance. In this context, hosts of machine learning techniques have demonstrated function approximation capabilities beyond regression methods. This study is aimed at comprehensive comparison of nine machine learning algorithms for developing a pavement performance model. Instrumental in pavement management systems, this extensive evaluation reconciles computational intelligence with systems engineering. Three principles of maximum likelihood, consistency, and parsimony are considered in algorithm benchmarking and model development. Along with quantitative measures of accuracy and generalization, qualitative criteria including scatter plots and sensitivity analysis are also implemented. Additionally, quantitative and qualitative aspects of model evaluation are combined through analysis of predicted performance trends. Results indicate that an algorithm performing the best in each comparison step might not achieve the same in the other. This comparison framework is recommended for identifying the appropriate machine learning paradigm to develop performance prediction models.
展开▼
